Testing Whether a One-Hand-Sized Mini PC Can Run a 27B Parameter AI Model

A palm-sized mini PC is put to the test on whether it can handle a serious 27 billion parameter AI model. The Khadas Mind Pro, the smallest Core Ultra X7 358H machine with Intel Panther Lake and 64GB of memory, is evaluated across four inference paths — CPU, Vulkan, SYCL, and OpenVINO — with the integrated Arc B390 GPU at the center of the measurements.

■ Hardware Overview
- Khadas Mind Pro specifications: Core Ultra X7 358H (Panther Lake), 64GB memory, integrated Arc B390 graphics
- The four available compute paths: CPU, Vulkan, SYCL, and OpenVINO

■ Benchmark Methodology
- Two measured phases: prompt processing (prefill/reading) and token generation (decode/writing)
- Bandwidth estimation derived from model size and measured token rate

■ Results & Analysis
- Prefill performance comparison across the four paths
- Decode performance range and how it relates to streaming model weights over shared system memory

Ideal for hardware enthusiasts and developers exploring local LLM inference on compact devices. It shows how measured performance differs across compute backends on integrated graphics and how memory bandwidth shapes token generation speed. Consider using the presented approach as a reference when evaluating mini PCs for on-device AI workloads.